StartLux-Decision offers models (0.8B to 27B) that use adaptive, confidence-guided reasoning via the RLCD^2 training pipeline. 'AutoThink' models trigger explicit reasoning only for uncertain cases, while 'Instant' models use bidirectional attention for fast, single-pass decisions.
HOW THIS AFFECTS YOU
●
builderYou can deploy models that balance fast inference with deep reasoning based on decision confidence.
●
researcherThe RLCD^2 pipeline uses proper scoring rules to reward improvements in decision probability quality.